Third Nomura Lecture From Dutch Dykes to Value-at-Risk ...embrecht/ftp/nomura28Apr04.pdf · { backtesting: RC based on kVaR10 ... † stress testing dependence: joint extremes, spillover,

Third Nomura Lecture

From Dutch Dykes to Value-at-Risk: Extreme ValueTheory and Copulae as Risk Management Tools

Paul Embrechts

ETH Zurich and London School of Economicswww.math.ethz.ch/~embrechts

University of Oxford, Nomura Centre for Quantitative Finance, April 27th, 2004

c©2004 (P. Embrechts, ETH Zurich)

En in alle gewesten

Wordt de stem van het water

Met zijn eeuwige rampen

Gevreesd en gehoord

(Marsman, 1938)

Throughout the centuries, the Dutch coast has been suffering numerous disastrous

floods:

1421, 1570, 1775, 1825, 1916, 1925, 1953, 2002*

1421: The St. Elisabeth flood

1570: The All Saints flood

* Not complete

c©2004 (P. Embrechts, ETH Zurich) 1

“Springtij en orkaan veroorzaken nationale ramp. Nederland in grote watersnood.”

(De Yssel - en Lekstreek, 6/2/53)

January 31 – February 1, 1953* (February flood):

• 1 836 people killed

• 72 000 people evacuated

• 49 000 houses and farms flooded

• 201 000 cattle drowned

• 500 km of sea dykes destroyed, ≥ 400 breaches

• 200 000 ha of land flooded

* Antwerp (Schoten), February 3, 1953


The Delta Project

• coastal defence against flooding

• required dyke height at location l: hδ(l)

• safety margin: MYSH(l)= Maximum Yearly Sea Height at location l:

P (MYSH(l) > hδ(l)) “small”

where “small” means:

* 10−4 at l = Randstad* 4−1 × 10−3 at l = Delta area and the North

• similar safety requirements for rivers but with safety values in the range of10−3–10−2

• solution for l = Randstad: hδ(l) = NAP + 5.14 m


Hence, we can reformulate the problem as:

• “small” = 1− α ∈ (0, 1) (small: EVT)

• l = fixed (different l: copulae)

• MYSH(l) = X, a rv with df F (co-variables: EVT, copulae, non-stationarity)

• hδ(l) = qα

qα = F −1(α) (α− quantile)

The February flood happened as a coincidence of several factors

• very low pressure over Scotland

• very strong North-Westerly storm

• “springtij” (extremely high tide)

All contributing to “the perfect storm scenario”


THE REGULATORY ENVIRONMENT

Financial Industry

• Basel Accord for Banking Supervision (1988)

– Cooke ratio, “haircut” principle, too coarse

• Amendment to the Accord (1996)

– VaR for Market Risk, Internal models, Derivatives, Netting

• Basel II (1998 – 2007)

– Three Pillar approach

– Increased granularity for Credit Risk

– Operational Risk


THE REGULATORY ENVIRONMENT

Insurance

• Solvency 1 (1997)

– Solvency margin as % of premium (non-life), of technical provisions (life)

• Solvency 2 (2000–2004)

– Principle-based (not rule-based)

– Mark-to-market (/model) for assets and liabilities (ALM)

– Target capital versus solvency capital

– Explicit modelling of dependencies and stress scenarios

• Integrated Risk Management


Value-at-Risk (Amendment to the Basel Accord, 1996)

Ingredients and Definition:

• V (τ): value of portfolio at time τ

• loss rv: L[τ,τ+∆t] = −(V (τ + ∆t)− V (τ))

• loss df: P(L[τ,τ+∆t] ≤ x

)(unconditional)

P(L[τ,τ+∆t] ≤ x|Ft

)(conditional)

(P & L: Profit and Loss Distribution)

often: τt = t×∆t, t ∈ N (∆t = 1 day, 10 days, 1 year)

• Lt+1 = L[τt,τt+1] = − (Vt+1 − Vt)


• the mapping: Vt = f (τt, Zt), Z ′t = (Z1,t, . . . , Zd,t)

(in practice, d is large, d ≥ 1 000, say)

• risk factor changes: Xt = Zt −Zt−1

• the loss operator maps risks into losses:

l[t](x) = − (f (τt+1, Zt + x)− f (τt, Zt)) , x ∈ Rd

henceLt+1 = l[t] (Xt+1)

• for f smooth (e.g. European Call case, Black-Scholes):

Lt+1 ≈ L∆t+1 := −

(fτ (τt, Zt)∆t +

d∑

i=1

fzi(τt, Zt) Xt+1,i

)

6

the Greeks (delta, “vega”, theta, rho)


Main approaches to (market) risk management:

(M1) Variance-covariance (normal) method

Assumption: Xt+1 ∼ Nd (µ, Σ)hence L∆

t+1 ∼ N1 (−ct− b′tµ, b′tΣbt)

(M2) Historical simulation

No assumptions!

Use empirical data: xt−n+1, . . . , xt and

construct pseudo portfolio values{

l̃s = l[t](xs) : s = t− n + 1, . . . , t}

yielding a histogram estimate of the P&L

(M3) Monte-Carlo simulation

Similar to (M2), but use historical data to calibrate some d−dimensional

model for the law of Xt+1 from which we simulate (Monte-Carlo) pseudo

portfolio values


How to measure risk?

ρ : M = {all portfolios} −→ R

satisfying certain axioms (ADEH-coherence) yielding for L ∈M:

ρ(L) = ρP(L) = sup{EQ(L) : Q ∈ P}

Examples in practice are:

• ρ1(L) = VaRα(L) = F−1L (α) (Lt+1 = L, say)

Value-at-Risk at confidence α (not always coherent)

• ρ2(L) = ESα(L) = E(L|L > VaRα(L))

Expected shortfall (easy adaptation is always coherent)


Basel Accord (Amendment) for Market Risk

• Internal models

• Regulatory capital based on 99% VaR with 10 day holding period

• Internal capital allocation using 95% VaR, 1 day

• Important issues:

– scaling:√

10−rule

– backtesting: RC based on k VaR1099, 3 ≤ k ≤ 4

– other risk categories (credit, operational)

– risk aggregation

Remark: Within Solvency 2 (Insurance) use ESα


FROM THE WORLD OF FINANCE

• “Extreme, synchronized rises and falls in financial markets occur infrequentlybut they do occur. The problem with the models is that they did not assign ahigh enough chance of occurrence to the scenario in which many things gowrong at the same time - the “perfect storm” scenario”

(Business Week, September 1998)

• Consulting for a large bank, topics to be discussed were:

– general introduction to the topic of EVT– common pitfalls and its application to financial risk management– the application of EVT to the quantification of operational risk– general introduction to the topic of copulae and their possible use in

financial risk management– sources of information to look at if we want to find out more

(London, March 2004)


EVT AND COPULAE

• Extreme Value Theory (EVT): stochastic models for

Mn = max(X1, . . . , Xn) (d = 1)

P (X − u ≤ x|X > u), u →∞ (d = 1, POT-method)

(non-obvious) generalization to d ≥ 2

• Copulae: multivariate df C on [0, 1]d with uniform marginals

FX(x) = P (X1 ≤ x1, . . . , Xd ≤ xd) , Xi ∼ Fi, i = 1, . . . , d

= C (F1(x1), . . . , Fd(xd))

FX ⇐⇒ (F1, . . . , Fd; C) (Sklar)

“⇒” : new copula models

“⇐” : stress testing


FOR THE QUANTITATIVE RISK MANAGER

WHY ARE COPULAE USEFUL

• pedagogical: “Thinking beyond linear correlation”

• stress testing dependence: joint extremes, spillover, contagion, ...

• worst case analysis under incomplete information:

given: Xi ∼ Fi, i = 1, . . . , d, marginal 1-period risks

Ψ(X): a financial position

∆: a 1-period risk or pricing measure

task: find min∆(Ψ(X)) and max∆(Ψ(X)) under the above constraints

• eventually: finding better fitting dynamic models


THE FUNDAMENTAL THEOREMS OF QUANTITATIVERISK MANAGEMENT (QRM)

• (FTQRM - 1) For elliptically distributed risk vectors, classical Risk

Management tools like VaR, Markowitz portfolio approach, ... work fine:

Recall:

– Y in Rd is spherical if Y d= UY for all orthogonal matrices U

– X = AY + b, A ∈ Rd×d, b ∈ Rd is called elliptical

– Let Z ∼ Nd(0,Σ), W ≥ 0, independent of Z, then

X = µ + WZ

is elliptical (multivariate Normal variance-mixtures)

– If one takes

W =√

ν/V , V ∼ χ2ν, then X is multivariate tν

W normal inverse Gaussian, then X is generalized hyperbolic


THE FUNDAMENTAL THEOREMS OF QUANTITATIVE

RISK MANAGEMENT (QRM)

• (FTQRM - 2) Much more important!

For non-elliptically distributed risk vectors, classical RM tools break down:

- VaR is typically non-subadditive

- risk capital allocation is non-consistent

- portfolio optimization is risk-measure dependent

- correlation based methods are insufficient

• A(n early) stylized fact:

In practice, portfolio risk factors typically are non-elliptical

Questions:

- are these deviations relevant, important

- what are tractable, non-elliptical models

- how to go from static (one-period) to dynamic (multi-period) RM


EXAMPLE: the Merton model for corporate default (firm value model,latent variable model)

• portfolio {(Xi, ki) : i = 1, . . . , d} firms, obligors

• obligor i defaults by end of year if Xi ≤ ki

(firm value is less than value of debt, properly defined)

• modelling joint default: P (X1 ≤ k1, . . . , Xd ≤ kd)

– classical Merton model: X ∼ Nd(µ, Σ)– KMV: calibrate ki via “distance to default” data– CreditMetrics: calibrate ki using average default probabilities for different

rating classes– Li model: Xi’s as survival times are assumed exponential and use Gaussian

copula

• hence standard industry models use Gaussian copula

• improvement using t-copula


• standardised equicorrelation (ρi = ρ = 0.038) matrix Σ calibrated so that fori = 1, . . . , d, P (Xi ≤ ki) = 0.005 (medium credit quality inKMV/CreditMetrics)

• set ν = 10 in t-model and perform 100 000 simulations on d = 10 000 obligorsto find the loss distribution

• use VaR concept to compare risks

Results:

min 25% med mean 75% 90% 95% maxGaussian 1 28 43 49.8 64 90 109 131

t 0 1 9 49.9 42 132 235 3 238

• more realistic t-model: block-t-copula (Lindskog, McNeil)

• has been used for banking and (re)insurance portfolios


Challenges (see www.math.ethz.ch/~embrechts)

EVT• extremes in economics and the economics of extremes

• EVT applications to operational risk

• non-stationarity

• multivariate EVT: P (X|X ∈ Hα) for P (X ∈ Hα) → 0

Copulae• dynamic models

• canonical limit theorems

• optimization results

• calibration within specific RM applications

• properties of statistical estimates


De meeste salicheyt hanght

aen de hoochte van eenen dijck

(Andries Vierlingh, Tractaet van Dyckagie, 1578)

Regulators have criticised LTCM and banks for not stress-testing risk

models against extreme market movements. ... The markets have

been through the financial equivalent of several Hurricane Andrews

hitting Florida all at once. Is the appropriate response to accept that

it was mere bad luck to run into such a rare event - or to get new

forecasting models that assume more storms in the future?

(The Economist, October 1998, after the LTCM rescue)


Third Nomura Lecture From Dutch Dykes to Value-at-Risk ...embrecht/ftp/nomura28Apr04.pdf · { backtesting: RC based on kVaR10 ... † stress testing dependence: joint extremes, spillover,

Documents